Valentyn Boreiko 🇺🇦
@valentynepii
Followers
173
Following
447
Media
10
Statuses
60
ML PhD student @ University of Tübingen from 🇺🇦. Working at the intersection of robustness, explainability and generative models. Ex-co-founder of Studyly.
Joined March 2022
RT @sbordt: Have you ever wondered whether a few times of data contamination really lead to benchmark overfitting?🤔 Then our latest paper a….
0
6
0
🚀Thrilled to share my ICML 2025 line-up!. Conference paper: “An Interpretable N-gram Perplexity Threat Model for LLM Jailbreaks” (Thu 17 Jul, 11:00–13:30 PDT, East Exhibition Hall A-B) - we propose a fluency-FLOPs threat model with adaptation of popular attacks to it. We can use.
0
1
8
Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.
488
237
3K
RT @maksym_andr: ⚠️Standard jailbreak attacks overfocus on info that can anyway be easily found online. However, LLM agents can cause much….
0
26
0
As jailbreaking attacks🤖on SOTA #LLMs (, LLM agents (, and even robots ( become more critical, our paper @ #NeurIPS2024 Red Teaming GenAI workshop proposes a unified threat model🛡️to compare these attacks!.
arxiv.org
The robustness of LLMs to jailbreak attacks, where users design prompts to circumvent safety measures and misuse model capabilities, has been studied primarily for LLMs acting as simple chatbots....
Ever wondered which jailbreak attack offers the best bang for your buck?🤖.Check out our #NeurIPS2024 workshop paper on a realistic threat model for LLM jailbreaks! We introduce a threat model that considers both fluency and .compute efficiency for fair attack comparisons. 🧵 1/n
0
1
7
RT @bfl_ml: Today we release the FLUX.1 suite of models that push the frontiers of text-to-image synthesis. read more at .
0
139
0
In our user study, distinguishing DVCs from real fundus images proved challenging, while spotting artefacts of the previous method (SVCs) was easy. More details in the paper, and you can explore the code here: (4/4).
github.com
Realistic retinal fundus and OCT counterfactuals using diffusion models and classifiers. - GitHub - berenslab/retinal_image_counterfactuals: Realistic retinal fundus and OCT counterfactuals using ...
0
0
1
We employ diffusion visual counterfactuals (DVCs) from to explain medical classifiers. These small image changes reveal insights into classifier predictions. (3/4)
1
0
2
See examples for fundus and OCT images when the starting images are of diseased or healthy classes. For fundus, we compare DVCs with artefacts introduced by a previous method (SVCs). Green arrows highlight changed diseased parts, yellow indicates SVC artefacts. (2/4)
1
0
1
Excited to share our collaborative work with @induilanchezian on explaining retinal fundus and OCT classifiers using diffusion models! 🧠👁️ Check out our latest paper: (1/4)
1
4
23
RT @evgenia_rusak: In our new paper (oral ICCV23), we develop a concept-specific pruning criterion (Density-Based-….
0
33
0
RT @chhaviyadav_: Join us 2mrw @⚡️XAI-in-Action : Past, Present & Future Applications⚡️ workshop @NeurIPSConf for an exciting discussion on….
0
11
0
RT @maksym_andr: 🚨 I'm looking for a postdoc position to start in Fall 2024!. My most recent research interests are related to understandin….
0
42
0
RT @NeuhausYannic: Check out our new paper “Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet”….
0
4
0
Joint work with @jan_metzen and Matthias Hein at @Bosch_AI and @uni_tue. Join us and learn more at #ICCV2023 “BRAVO” workshop in Paris on October 2nd from 15:30 - 16:15! See also our paper under .
arxiv.org
The identification and removal of systematic errors in object detectors can be a prerequisite for their deployment in safety-critical applications like automated driving and robotics. Such...
0
0
2
It is also possible to quantitatively evaluate systematic error found in such a way using average error rate. For example, when we leave a color background for simplicity, we can see the following systematic errors for different object detectors.
1
0
0
SCROD supports tests on realistic backgrounds too. For example, FCOS predicts “suitcase” for the images of a coupe car of color purple at a particular scale, rotation angle (80 degrees) (LEFT). However, by changing the rotation angle to 70 degrees (RIGHT), it predicted “car”.
1
0
0
SCROD however can generate images for the same setting, but with full control and can discover systematic errors of object detectors. FasterRCNN2, for example, detects a sports car of blue color at a particular scale, rotation angle, and with a gray background as an “airplane”.
1
0
0
A single generative model cannot give full control over the generated scene. SCROD addresses it for the street scene synthesis. SD-v1.5 for the prompt “blue sports car rotated by 50 degrees around the X axis from the side view on a gray background in the center of the image”:
1
0
0
Object detectors are important in many applications such as autonomous driving - that’s why we propose “Segment Control Rotate Outpaint Detect” (SCROD) - a pipeline of generative models that automatically searches for hard cases for object detectors that persist over 16 seeds.
1
0
0
Excited to share our #ICCV2023 BRAVO workshop paper on the SCROD pipeline!.SCROD allows fine-granular control of object pose and appearance and by this is able to identify systematic errors of object detectors in rare situations such as the one shown below.
2
4
15